allocating systematic and unsystematic risks in a ... · pdf fileto disentangle the systematic...

38
Allocating Systematic and Unsystematic Risks in a Regulatory Perspective C., GOURIEROUX (1) and A., MONFORT (2) (First version, September 2010, revised September, 2011) The first author gratefully acknowledges financial support of the NSERC Canada and the chair AXA/Risk Foundation : ”Large Risks in Insurance”. We thank M., Summer, D., Tasche for valuable feedback. The views ex- pressed in this paper are those of the authors and do not necessarily reflect those of the Banque de France. 1 CREST, and University of Toronto. 2 CREST, Banque de France, and University of Maastricht. 1

Upload: vantuong

Post on 27-Mar-2018

218 views

Category:

Documents


2 download

TRANSCRIPT

Allocating Systematic and Unsystematic Risksin a Regulatory Perspective

C., GOURIEROUX (1) and A., MONFORT(2)

(First version, September 2010, revised September, 2011)

The first author gratefully acknowledges financial support of the NSERCCanada and the chair AXA/Risk Foundation : ”Large Risks in Insurance”.We thank M., Summer, D., Tasche for valuable feedback. The views ex-pressed in this paper are those of the authors and do not necessarily reflectthose of the Banque de France.

1CREST, and University of Toronto.2CREST, Banque de France, and University of Maastricht.

1

Allocating Systematic and Unsystematic Risks in a RegulatoryPerspective

Abstract

This paper discusses the contributions of financial entities to a globalreserve from a regulatory perspective. We introduce axioms of decentral-ization, additivity and compatibility with risk ordering, which should besatisfied by the contributions of the entities and we characterize the set ofcoherent contributions compatible with these axioms. Then, we explain howto disentangle the systematic and unsystematic risk components of these con-tributions. Finally, we discuss the usual relationship between baseline reserveand reglementary required capital, and propose alternative solutions to thequestion of procyclical required capital.

Keywords : Risk Measure, Allocation, Regulation, Coherent Contribution,Systemic Risk, Procyclical Effect.

1

1 Introduction

The definition of the required capital in Basel regulation is often presentedas an important reason of the development of the recent financial crisis. Thefollowing arguments are in particular invoked :

i) The regulatory capitals, which the banks were required to hold by theregulator, were not sufficiently large to cover the (extreme) risks.

ii) They did not account for comovement of financial institutions assets andliabilities, that is, for systematic risk factors.

iii) This regulation had a procyclical effect, instead of the expected counter-cyclical effect.

These possible drawbacks of the previous regulation explain the recentchanges in both regulation and academic research. Examples are the in-troduction of additional regulators focusing on systemic risk, the variousstress-testing performed in US as well as in European countries, or the newmeasures of systemic risk introduced in the academic literature.

The aim of our paper is to identify more precisely possible deficiencies ofthe regulation rules and to propose alternative methods. Let us first recallhow Basel regulation defines the required capital of a given entity. In a firststep each financial entity has to compute a measure of its own risk (and alsosuch measures for each business line separately.) The standard risk measureused by these entities is the Value-at-Risk (VaR), which gives the maximumloss within a α% confidence interval. The level α is fixed by the regulator andthe risk of an individual entity is considered in isolation. Then, in a secondstep, the required capital is fixed from the observed individual history ofVaR. A typical formula for required capital at day t is for instance :

RCt = max(V aRt, k1

60

59∑

h=0

V aRt−h), (1.1)

where V aRt denotes the VaR at horizon 1-year and the trigger parameterk depends on the technical level of the entity and is generally larger than3. This nonlinear link function between the risk measure and the requiredcapital induces two regimes : in a standard risk environment for the bank,

2

formula (1.1) reduces to : RCt = k1

60

59∑

h=0

V aRt−h. The smoothing of risk

measures over 60 opening days, i.e. 3 months, is introduced in order to avoidan erratic evolution of the reserves. Trigger coefficient k provides an addi-tional insurance against risk. When the entity becomes suddenly very risky,

that is, when the current risk measure V aRt is larger thank

60

59∑

h=0

V aRt−h,

the required capital becomes equal to V aRt.

The potential drawbacks of this kind of approach are threefold. First itseems questionable to consider each entity separately without any referenceto its role in the global system. Our approach is a top-down method focusingon the way of defining the contributions of the entities to a global reserveneeded for the whole system. The second drawback of the standard approachis the lack of distinction between systematic and unsystematic componentsof the risk, which will be separated in our approach. Third, the usual methodmay imply procyclicity that we try to avoid in our framework.

In order to deal with these problems we first consider general axiomswhich should be satisfied by the set of contributions to a global risk in anyparticular context. We thus obtain characterizations of ”coherent” contri-butions in an agnostic or economically neutral way. Then we show howsuch contributions can be decomposed into a systematic and an unsystem-atic part and we identify the degrees of freedom remaining at the disposal ofthe decision makers in a specific context. The non-uniqueness of the coher-ent contributions is a good thing since it would be unreasonable to expect touniquely determine them only by universal axioms.

The paper is organized as follows. In Section 2, we explain how to allocatethe global reserve among the entities. We introduce the decentralization, ad-ditivity and risk ordering axioms, which are relevant for this decomposition,and characterize the contributions satisfying the three axioms. In Section3, we discuss the alternatives proposed in the literature to define the contri-butions from the regulatory perspective, such as the CoVar, Shapley values,or Euler allocation 3. In Section 4 we derive a disaggregation formula not

3The research of an appropriate allocation of capital for a purpose internal to a bank ,for instance to maximize shareholder value, achieve capital efficiency, or measure concen-

3

only in terms of entities, but also in terms of systematic and unsystematicrisks, both in linear and nonlinear factor models. Section 5 explores the linkbetween the required capital and the objective measures of systematic andunsystematic risks. In particular, we explain why the Through The Cycle(TTC) smoothing treatment of these components have to be performed sepa-rately to avoid the spurious procyclical effect of the standard regulation andwe propose alternative solutions to avoid this procyclicity of the requiredcapital. Section 6 concludes. Technical proofs are gathered in appendices.

2 Disaggregation of a global reserve

Let us denote by Xi, i = 1, . . . , n, the Loss and Profit (L&P) of the entities(banks). The global L&P, that is the L&P of the whole system, for instance

the banking system, is : X =n∑

i=1

Xi. The global reserve for the system, that

is the reserve for systemic risk, is R(X). It is assumed to depend on thedistribution P of the global L&P only. At this point, we do not discuss howthe global reserve is computed. It can be a VaR, or a coherent risk measure,such an Expected Shortfall (ES), or something else (see Section 5). Thisglobal reserve has to be assigned to the different entities :

R(X) =n∑

i=1

R(X,Xi), (2.1)

say, where R(X,Xi) denotes the contribution of entity i to the total reserve.Moreover, we would like to decompose the individual contribution into :

R(X,Xi) = Rs(X,Xi) +Ru(X,Xi) +Rs,u(X,Xi), (2.2)

where Rs (resp. Ru) denotes the contribution for marginal systematic (resp.unsystematic) risk and Rs,u the contribution for the cross effects. At this levelnote that the different components of the contribution might take negativevalues. This point has already been mentioned when the reserves for Basel1-2 were computed by a VaR, or an expected shortfall, even if such negativevalues have rarely been observed in practice. A negative value of a VaR

tration risk in a portfolio [see e.g. Patrick et alii (1999), Dhaene, Goovaerts, Kaas (2003),Sherris (2007), Tasche (2008)] is clearly out of the scope of the present paper.

4

arises if a financial institution is too risk averse; in this case, this is anincentive to invest in risky assets and indirectly to improve the liquidityof some markets for risky assets without increasing to much the risk. Thepossibility of a negative value for the reserve for systemic risk has a differentpurpose. It would imply a total reserve smaller than the standard one, butnot in general a negative total reserve for the institution. Such a situationmight occur more frequently, and is well illustrated by a simple example. Letus consider a system with an institution with a portfolio hedging the completebanking system against systemic risk, called below insurance company, andseveral banks with very risky investments with respect to this systemic risk.If the reserves for systemic risk are constrained to be nonnegative, they willbe equal to zero for the insurance company. In other words the hedgingagainst the risk on the banking system due to this company is priced at zerovalue. This is a public good the other risky institutions can profit of for free.With the possibility of negative reserve for systematic risk, we introduce avalue for this hedging, and implicitly a payment of this partial systematicrisk insurance by the risky institutions without introducing a new market forsystemic risk. Compared to the previous Basel 2 regulation, the introductionof reserves for systemic risk is more an adjustment, a kind of systemic oriented”bonus-malus” on the reserves.

We focus in this section and the following one on decomposition (2.1) anddefer the main discussion on systematic risk to Sections 4 and 5.

2.1 Coherent contributions

From an axiomatic point of view, it is important to distinguish the mea-sure of the regulatory capital for the global risk, that is the function R(.) :X → R(X), and the contributions to the total reserve, that is, the functionR(X, .) : Xi → R(X,Xi). These contributions are contingent to the totalrisk level and should satisfy at least a decentralization axiom, an additivityaxiom and a risk ordering axiom.

i) Decentralization axiom

A1. Decentralization axiom : The individual contribution R(X,Xi)of entity i depends on the joint distribution of (X,Xi), but is independentof the decomposition of X −Xi into Σj 6=iXj.

5

This axiom has been first introduced in Kalkbrener (2005). In a reg-ulatory perspective, it has the advantage of allowing for a computation ofR(X,Xi) by entity i, while preserving a minimal confidentiality on the indi-vidual portfolios of the other entities.

More precisely, let us consider entities invested in stocks. The individualL&P ′s are : Xi = Y ′γi, where Y denotes the vector of share values of thestocks, and γi is the portfolio composition of the loss for entity i. The reservesare usually evaluated for a crystallized portfolio, that is, with the compositionγ−i existing at the beginning of period t (end of period t−1). With this prac-

tice, the regulator has to provide entity i with the type of measure R(X,Xi)

to consider, the past data on Y , and the sumn∑

j=1

γ−j corresponding to the

global crystallized portfolio, without providing the individual information oncompetitors’portfolios γ−

j ,∀j 6= i.

ii) Additivity axiom

The additivity axiom has been first introduced in Garman (1997) [seealso Kalkbrener (2005), where it is called linear aggregation axiom].

A2. Additivity axiom :

R(X) =n∑

i=1

R(X,Xi), for any decomposition of X into X =n∑

i=1

Xi.

The total reserve should not depend on the number of entities holdingthe risk and of their respective sizes, whenever the sum of these L&P ′s staysthe same. The additivity axiom has several consequences.

i) For n = 1, we get : R(X) = R(X,X).ii) We also have :R(X) = R(X,X1 +X2) +R(X,X −X1 −X2)

= R(X,X1) +R(X,X2) +R(X,X −X1 −X2),which implies :

R(X,X1 +X2) = R(X,X1) +R(X,X2),∀X1, X2. (2.3)

This is the additivity property of the function R(X, .) : Xi → R(X,Xi),for any given X. By imposing the additivity axiom, any merging, or demerg-

6

ing of entities without effect on global risk provides no spurious advantagein terms of contribution to the total reserve. It is introduced to avoid thatsome institutions circumvent the spirit of the regulations for systemic risk byeither merging, or demerging.

Note that the standard stand-alone risk measures such as the V aR(Xi) orany other coherent risk measure cannot be chosen for defining contributions.Indeed, they are subadditive [Artzner et al. (1999), Acerbi, Tasche (2002)],and thus do not satisfy the axiom of additivity. The main reason is thatthe standard measures of individual risk such as the VaR and the expectedshortfall are not contingent to the level of global risk. According to Tasche(2008) (Remark 17.2), the choice of stand-alone risk measure values as riskcontributions would more the banks which improve the diversification of theglobal regulatory portfolio.

iii) Risk ordering axiom

The contributions have also to be compatible with an appropriate notionof stochastic dominance. Intuitively, the contribution has to take into ac-count not only the individual risk of entity i, but also its hedging potentialwith respect to the set of other entities. Thus, we have to introduce a di-rectional notion of stochastic dominance valid for an individual risk X1, say,and a given global L&P : X. For this purpose, let us consider the virtualdecomposition of the global portfolio into X1 and X2 = X −X1 obtained byaggregating the L&P of the other entities.

Definition 2.1 : Let us consider the L&P : X,X1, X∗1 . We say that X∗

1

stochastically dominates X1 at order 2 with respect to X, if and only if :

E[U(X∗1 , X −X∗

1 )|X] ≥ E[U(X1, X −X1)|X],

for any concave function U .

This is a second-order stochastic dominance [Rothschild, Stiglitz (1970)],applied to virtual portfolios X1, X2, whose allocations are constrained to sumup to a given X. The directional stochastic partial ordering is denoted byºX .

The directional stochastic dominance can be characterized in simpler ways

7

(see Appendix 1).

Proposition 2.2 : We have the following equivalences :i) X∗

1 ºX X1;ii) E[U(X∗

1 )|X] ≥ E[U(X1)|X], for any concave function U ;iii) There exists a variable Z such that :

X1 = X∗1 + Z, with E(Z|X,X∗

1 ) = 0.

Proposition 2.2 shows that the directional stochastic dominance is equiv-alent to the standard second-order stochastic dominance applied to the con-ditional distribution of X1 given X [see Rothschild, Stiglitz (1970)].

The next axiom concerns the compatibility of the contribution with thedirectional stochastic dominance.

A.3. Risk ordering axiom

We have R(X,X∗1 ) ≤ R(X,X1) for any pair of entity risks such that

X∗1 ºX X1.

The next proposition characterizes the contributions satisfying the de-centralization, additivity and risk ordering axioms, called coherent contribu-tions.Proposition 2.4 : Let us assume that the L&P : X,Xi belong to a spaceL2(Y ), where Y is a given set of random variables. If the contribution func-tion R(X, .) : Xi → R(X,Xi) is continuous with respect to Xi for the L2-norm, then, the coherent contributions satisfying Axioms A1, A2, A3, are

such that RµP(X,Xi) =

∫E(Xi|X = x)µP (dx), where µP is a measure (not

necessarily a probability measure), which can depend on the distribution P

of X and is such that

∫xµP (dx) = R(X).

Proof : See Appendix 2,

The conditionXi ∈ L2(Y ) means that the portfolios of interest are writtenon some basic assets (with prices Y ), possibly including derivatives withnonlinear (square integrable) payoffs.

8

When the distribution of X is continuous, with a strictly increasing cu-mulative distribution function (cdf), the contributions in Proposition 2.4 canbe written in terms of quantiles.

Corollary 2.5 : The coherent contributions are also characterized by :

RνP (X,Xi) =

∫E[Xi|X = qα(X)]νP (dα),

where νP is a measure, which can depend on the distribution P of X and issuch that :

∫qα(X)νP (dα) = R(X).

Proof : The result is obtained by applying the change of variable x = qα(X).

QED

The equality R(X) =

∫qα(X)νP (dα) implies that R(X) is a weighted

quantile. νP looks like a usually distortion measure, except that it can dependon the distribution of X, since the contribution is contingent to X. It will becalled the allocation distortion measure (ADM) in the rest of the paper.

More precisely, let us assume for illustrative purpose that the global riskmeasure is equal to a distortion risk measure (DRM) that is, a weightedcombination of VaR’s [see Wang (2000), Acerbi (2002)] :

R(X) = DRMH(X) =

∫qα∗(X)H(dα∗),

where H denotes a distortion (or spectral) probability measure on (0, 1).This measure H is usually fixed independently of the distribution P of theglobal risk; for instance it is equal to a uniform distribution when R(X) isthe expected shortfall. Corollary 2.5 says that we do not have necessarilyto use the fixed distortion measure H defining the DRM as the allocationdistortion risk measure. Moreover, a given level of global reserve, 2 billions$, say, can be seen as the value of a VaR as well as the value of an ES4,

4Indeed the functions α → V aRα and α → ESα are continuous increasing functions ofthe critical value α; therefore, there exist two critical values α1, α2 such that :

V aRα1(X) = ESα2(X) = 2 billions $.

9

and more generally as the value of an infinite number of alternative DRM,

whenever R(X) =

∫qα(X)νP (dα).

As for the notion of coherent risk measure [see Artzner et al. (1999)], theaxiomatic of coherent contributions restrict the set of possibilities to easilyinterpretable allocations, but does not provide a unique solution.

Corollary 2.6 : The contributions defined in Proposition 2.4 do not dependon the measure µP such that

∫xµP (dx) = R(X) when (X,X1, . . . , Xn) is

Gaussian and are all equal to E(Xi) +Cov(Xi, X)

V (X)[R(X)− E(X)].

Proof : It is a consequence of the formula E[Xi|X = x] = E(Xi) +Cov(Xi, X)

V (X)(x−E(X)), which impliesRµP

(X,Xi) = E(Xi)+Cov(Xi, X)

V (X)[R(X)−

E(X)], by integrating w.r.t. µP .

QED

Corollary 2.7 : If R(X,Xi) is a coherent contribution and if X∗i ºX Xi, we

have R(X,Xi) = R(X,X∗i ).

Proof : From Proposition 2.2 iii), we get E(Xi|X = x) = E(X∗i |X = x) and

the result follows from Proposition 2.4.

QED

Corollary 2.7 shows the difference between a marginal (or unconditional)measure of risk (such as a VaR) and a contribution. Changing X∗

i intoX∗i +Z

increases the marginal risk in the second-order dominance sense. However,due to the decentralization and additivity axioms this increase is not takeninto account in the contribution because of the compensation between X∗

i

and the other L&P ′s, if X stays the same.

3 Related literature

There exist three streams of literature for defining risk measures contingentto the level of global risk. Some authors propose alternative sets of axioms to

10

be satisfied by the contributions. Other authors focus on the direct introduc-tion of contingent reserve levels R(X,Xi). The CoVaR, or the computationof the reserve based on the Shapley value belong to this literature. Finally,it is possible to directly introduce a decomposition formula (2.1) of the totalreserve, and try to interpret ex-post the elements R(X,Xi) in this decom-position. The Euler allocation provides an example of this approach. Let usbriefly review these approaches. 5

3.1 Alternative sets of axioms

Other sets of axioms have been considered in the literature. For instanceKalkbrener (2005) [see also Hesselager, Andersson (2002), Furman, Zitikis(2008) for a similar approach] proved the uniqueness of the contributionunder the linearity assumption, an additional continuity assumption and thefact that :

Diversification axiom :

R(X,Xi) ≤ R(Xi, Xi) = R(Xi),∀X,Xi.

We do not introduce this diversification axiom in our approach. Indeedthe axiom involves two different levels of global risk, and thus introducejoint restrictions between the allocation function Xi → R(X,Xi) for fixedX, and the global reserve function X → R(X). This explains one of theconsequence of the set of axioms by Kalkbrener (2005), which is the positivehomogeneity and subadditivity of the global measure R(.). As mentionedearlier and also discussed in Section 5.1, these latter features of function R(.)are not appropriate in a regulatory perspective.

3.2 The CoVaR

Adrian and Brunnermeier (2009) propose to analyze the risk of the entities,when the system is in distress. More precisely, let us denote by qα(X) the α-quantile corresponding to the system6. The CoVaR for entity i and confidencelevel α when the system is in distress is defined by :

5All these approaches might also lead to negative contributions for systemic risk.6A similar approach can be based on another risk measure such as the Expected Short-

fall [see e.g. Kim (2010)].

11

P [Xi < CoV aRi|s,α(X)|X = qα(X)] = α. (3.1)

The CoVar is generally larger than the VaR of the entity. It is implicitlyproposed to choose their difference, called ∆CoVar, as the contribution tosystematic risk :

Rs(X,Xi) = CoV aRi|s,α(X)− qα(Xi), (3.2)

with Ru(X,Xi) = qα(Xi). In this case the contribution of entity i to the totalreserve i is :

R(X,Xi) = CoV aRi|s,α(X).

The ∆CoVar is an interesting measure of systematic risk, but the Co-VaR itself cannot be used directly for defining the contribution. Even ifthe decentralization axiom is satisfied, the CoVaR as the VaR based ap-proaches are bottom-up approaches. In particular, the associated total re-

serven∑

i=1

CoV aRi|s,α(X) is not a function of the distribution of the total risk

only and the additivity axiom is not satisfied.

3.3 The Shapley value

A Shapley value [Shapley (1953)] is a fair allocation of gains obtained bycooperation among several actors. Let us assume that all actors i = 1, . . . , naccept to cooperate and introduce a superadditive value function v(S), whichmeasures the gain of this cooperation for a coalition S ⊂ {1, . . . , n}. Thesuperadditivity condition :

v(S ∪ T ) ≥ v(S) + v(T ),

expresses the fact that cooperation can only be profitable.The Shapley value is one way to distribute the total gains of the players,

if they all collaborate, by defining an added value v(S ∪ {i}) − v(S) whenactor i joins coalition S. The Shapley value is defined as a mean of thesequantities over all possible coalitions :

Vi =∑

S⊂{1,...,n}\{i}

{ |S|!(n− |S| − 1)!

n![v(S ∪ {i})− v(S)]

}, (3.3)

12

where |S| denotes the number of actors in coalition S. Denault (2001),Koyluoglu, Stocker (2002), Tarashev et alii (2009) (with the Varying TailEvents procedure) propose the Shapley value as a fair allocation of the reservewith v(S) = −R(Σi∈SXi), and R(.) a risk measure such as a VaR, or anexpected shortfall.

In a regulatory perspective, the drawback of this approach is twofold :

(*) It assumes a total cooperation of the entities, which are in practicecompetitors.

(**) The Shapley allocation does not satisfy the decentralization, i.e. con-fidentiality, axiom which intuitively is not compatible with a total co-operation.

In a regulatory perspective, such a Shapley allocation could only be im-plemented by the regulator itself due to the confidentiality restriction. Thiswould lead to a highly centralized computation of the contributions by theregulator itself, but is clearly not implementable in practice. First, the regu-lator does not possess the technical departments to make such computationsfor all entities. Second, such a centralized approach contradicts the spirit ofthe second Pillar of Basel 2 regulation, where the entities have to learn howto manage and control their internal risks by themselves.

3.4 Euler allocation

As noted in Litterman (1996), p28, and Garman (1997), footnote 2, if thefunction R(.) defining the total reserve is homogenous of degree 1, that is,satisfies the condition :

R(λX) = λR(X), ∀λ > 0, (3.4)

or, equivalently, R∗(λe) = λR∗(e),∀λ > 0,

where R∗(λ1, . . . , λn) = R(n∑

i=1

λiXi), we get the Euler condition :

R∗(e) =n∑

i=1

∂R∗(e)∂λi

, (3.5)

13

obtained by differentiating both sides of equation (3.4) with respect to λ.This provides a decomposition of the total reserve as the sum of its sensi-tivities corresponding to shocks performed separately on each entity. Thisjustifies the terminology Euler allocation used in McNeil et al. (2005), Sec-tion 6.3.

Let us consider this decomposition for a global reserve defined by a dis-tortion risk measure :

R(X) = DRMH(X) =

∫qα(X)H(dα), (3.6)

where H denotes a distortion (probability) measure on (0,1).

The VaR and more generally any DRM is homogenous function of degree1. Thus, the total reserve can be decomposed into :

DRMH(X) =n∑

i=1

DRMH,i, (3.7)

where the marginal expected distortion risk measures are given by :

DRMH,i =

∫qα,iH(dα), (3.8)

with qα,i =∂q∗α(e)∂λi

and q∗α(λ1, . . . , λn)′ = qα(

n∑i=1

λiXi), . (3.9)

The following result has been derived in Gourieroux, Laurent, Scaillet(2000) (see also the beginning of Appendix 3, formula a.3).

Proposition 3.1 :

qα,i =∂q∗α(e)∂λi

= E[Xi|X = qα(X)].

and

DRMH,i =

∫E[Xi|X = qα(X)]H(dα).

This Euler allocation applied to a DRM satisfies the three axioms of Section2.1. However, it is rather restrictive, since it assumes that function R(.)

14

is a DRM, corresponds to the choice of an allocation distortionmeasure equal to the distortion measure itself.

The Euler decomposition formula (3.7) is in particular valid for the ex-pected shortfall, for which the distortion measure is the uniform distributionon the interval (α, 1) [Wang (2000), Acerbi, Tasche (2002)] :

ESα(X) =1

1− α

∫ 1

α

qα∗(X)dα∗. (3.10)

It can be checked that the marginal expected shortfall is equal to :

MESi =1

1− α

∫ 1

α

E[Xi|X = qα∗(X)]dα∗

= E[Xi|X > qα(X)]. (3.11)

This simplified expression of the marginal expected shortfall has beenfirst derived by Tasche (2000) [see also Kurth, Tasche (2003) and Appendix3].

The Euler decomposition of the VaR and the Expected Shortfall abovediffer essentially by the conditioning set. They stress that the additive de-compositions involve both conditioning with respect to system distress [asin definition (3.4) of the CoVaR] and conditional expectations (instead ofconditional quantiles as in the CoVaR approach) to ensure the additivityproperty.

However, when R(.) is assumed to be a VaR, or an expected shortfall,R(.) is homogenous of degree one and this homogeneity assumption of theglobal risk measure R(.) is questionable, especially if the regulation is usedfor economic policy. As an illustration, let us assume that the portfolios ofinterest include the different types of credits. From a macroeconomic pointof view, there exists an optimal level for the global amount of credit to bedistributed in the economy. The global risk measure has to be chosen as anincentive to reach this optimal level. The cost of the reserve has to be small,if the current amount of credit is below this optimal level, large, otherwise.Mathematically, we expect function λ → R(λX)/λ to be increasing in λ, notconstant. For instance Rc(X) = E(X)+E[(X−c)+], where c is the ”optimallevel” would satisfy this condition7.

7The homogeneity assumption is sometimes justified by an invariance of the risk mea-

15

It is important to note that the condition on function, R(.) assumed inCorollary 2.5 does not imply the homogeneity of degree 1, since the ADMcan vary with the distribution of systemic risk X. More precisely, we have :

R(λX) =

∫qα(λX)νPλ

(dα)

= λ

∫qα(X)νPλ

(dα)

6= λ

∫qα(X)νP1(dα) = λR(X),

since the ADM can depend on the distribution of X and thus on λ.

4 Contribution to systematic risk

The axiomatic approach of Section 2, or the Euler allocation of Section 3.4provide contributions of individual entities satisfying the axioms, but do nottry in general to reallocate the global systemic risk between its systematicand unsystematic components. The aim of this section is to explain howthe allocation principle can be applied to disentangle the systematic andunsystematic components of the systemic risk. For expository purpose, wefirst consider models with linear factors driving the systematic risk, whichare usually considered when we focus on market risk. Then, the approach isextended to nonlinear factors. Nonlinear factor models are involved wheneveroptions and/or credit risks are considered.

4.1 Linear factor model

Let us first consider a linear factor model. The individual L&P ′s can bedecomposed as :

Xi =K∑

k=1

βikfk + γiui, (4.1)

sure with respect to a change of money unit. Function Rc(X) satisfies this latter conditionwhich involves the changes X → λX and c → λc, since the optimal level is also written inmoney unit, but is not homogenous of degree 1 in X for fixed c.

16

where f1, . . . , fK are systematic factors and u1, . . . , un idiosyncratic termswith K < n. These factors are random at the beginning of period t, andobserved at the end of this period. For instance, for fixed income derivatives,the main risk factors can be the interest rate level, slope and curvature,the spreads over T-bond rates, the exchange rates... The total L&P can bedecomposed as :

X =K∑

k=1

βkfk +n∑

i=1

γiui, (4.2)

with βk =n∑

i=1

βik.

i) Euler allocation of a VaR global risk measure

For expository purpose, let us first assume that the global risk measureis a VaR : R(X) = qα(X).

The α-quantile of X is a function :

qα(X) = q∗α (β1, . . . , βK , γ1, . . . , γn, θ) , (4.3)

where θ denotes the parameters characterizing the joint distribution off1, . . . , fK , u1 . . . , un.

The marginal effect of an homothetic change of exposure of the entities

passing from Xi to λXi =K∑

k=1

λβikfk + λγiui, i = 1, . . . , n gives :

qα(X) =

[dqα(λX)

]

λ=1

=K∑

k=1

(n∑

i=1

βik)∂qα(X)

∂βk

+n∑

i=1

γi∂qα(X)

∂γi(4.4)

=n∑

i=1

qα,i.

where the composite term :

17

qα,i ≡K∑

k=1

βik∂qα(X)

∂βk

+ γi∂qα(X)

∂γi, (4.5)

shows how the VaR Euler contribution qα,i of entity i can be decomposed inorder to highlight the effects of systematic factors and idiosyncratic term.

Decomposition formulas (4.4)-(4.5) explain how to pass from Euler alloca-tions computed by entity to Euler allocations computed by ”virtual businesslines” associated with the different risk factors, as summarized in Table 1.In other words, we propose to treat in a symmetric way the contributions tosystemic risk of both risk factors and entities.

18

Table 1 : Euler decomposition of a global VaR

entity 1 i n risk contributionrisk factor of the factors

f1...

fk βik∂qα(X)

∂βk

βk∂qα(X)

∂βk...fKu1...

ui 0 γi∂qα(X)

∂γi0 γi

∂qα(X)

∂γi...un

risk contribution qα,i V aR(X)of the entities

By applying the expression of the VaR sensitivity (see Proposition 3.1),we get as a by-product the Euler components associated with systematic andunsystematic risks, respectively, as :

R(X) = Rs(X) +Ru(X),

with Rs(X) =n∑

i=1

Rs(X,Xi), Ru(X) =n∑

i=1

Ru(X,Xi),

Rs(X,Xi) =K∑

k=1

βik∂qα∂βk

(X) =K∑

k=1

βikE[fk|X = qα(X)], (4.6)

Ru(X,Xi) = γi∂qα∂γi

(X) = γiE[ui|X = qα(X)], (4.7)

ii) General case

19

The example of Euler allocation of a VaR discussed in Table 1 providesa principle of allocation between systematic and unsystematic risks valid ina general framework where function R(.) is not necessarily a DRM, or evenan homothetic function. Indeed, whenever we have R(X) =

∫qα(X)νP (dα),

we can use the contribution derived in Corollary 2.5, which takes a formsimilar to a contribution associated with an Euler allocation. The idea isto define the new entities by crossing the initial entity and the type of risk.The L&P for entity i and systematic risk (resp. unsystematic risk) is :

Xs,i =K∑

k=1

βikfk(resp. Xu,i = γiui). The components of the total L&P are

defined accordingly by : Xs =n∑

i=1

Xs,i and Xu =n∑

k=1

Xu,i. Then, for a given

ADM νP , the allocations are defined by :

RνP,s(X,Xi) = RνP (X,Xs,i),

=K∑

k=1

βikRνP (X, fk)

RνP,u(X,Xi) = RνP (X,Xu,i),

= γiRνP (X, ui)

RνP (X,Xi) = RνP,s(X,Xi) +RνP,u

(X,Xi),

RνP,s(X) =

n∑i=1

RνP,s(X,Xi) = RνP (X,Xs),

RνP,u(X) =

n∑i=1

RνP,u(X,Xi) = RνP (X,Xu).

with RνP (X, .) =

∫E[.|X = qα(X)]νP (dα) (see Corollary 2.5).

These contributions correspond to the axiomatic of Section 3 applied todepartments specialized in systematic (resp. unsystematic) risk components.

20

It is justified whenever there exist on the market financial products intro-duced to replicate the systematic factor.

4.2 Nonlinear factor model

The allocation of the global reserve among the entities can be done for bothlinear and nonlinear factor models. However, the allocation between system-atic and unsystematic components is less than obvious if :

Xi = gi(f, ui), (4.8)

where the (multidimensional) factor f and the idiosyncratic term ui are in-dependent, but function gi is nonlinear, due to the presence of cross effects.

Nevertheless, it is possible to decompose the individual L&P as :

Xi = E(Xi|f) + [E(Xi|ui)− E(Xi)] + [Xi − E(Xi|f)− E[Xi|ui] + E(Xi)]

≡ Xs,i +Xu,i +Xu,s,i, say, (4.9)

where Xs,i, Xu,i, Xs,u,i are the marginal systematic and unsystematic effects,and the cross effect, respectively 8. Even if the systematic factor f and theidiosyncratic terms ui are independent, interaction effects will appear in therisk contributions due to the nonadditive decomposition (4.8).

Then, the total contribution of entity i can be decomposed as :

RνP (X,Xi) = RνP ,s(X,Xi) +RνP,u(X,Xi) +RνP,s,u

(X,Xi), (4.10)

where :

with RνP,s(X,Xi) = RνP (X,Xs,i)

RνP,u(X,Xi) = RνP (X,Xu,i)

RνP,s,u(X,Xi) = RνP (X,Xu,s,i)

and Rνp(X, .) =

∫E[.|X = qα(X)]νP (dα)

8This type of decomposition can also be used to distinguish the effects of dependentsystematic factors [see Rosen, Saunders (2010), Section 4.5]

21

This decomposition does not depend on the selected representations ofthe factor and the idiosyncratic term, that is, the decomposition is invariantwhen either f , or ui is transformed by a one-to-one transformation.

As an illustration, let us consider a model with stochastic drift and volatil-ity driven by a same factor f :

Xi = mi(f) + σi(f)ui, i = 1, . . . , n,

where f is independent of u = (u1, . . . , un)′, and the errors are iid zero mean.

We have, using R(X) = qα(X) :

Rs(X,Xi) = E[mi(f)|X = qα(X)],

Ru(X,Xi) = E[E{σi(f)}ui|X = qα(X)],

Rs,u(X,Xi) = E[(σi(f)− E{σi(f)})ui|X = qα(X)].

This example shows that in a nonlinear model, the effect of the systematicfactor is captured by both Rs and Rs,u. Their relative magnitude can behighly different for the different entities. For instance in a basic stochasticvolatility model, where mi(f) = 0, only the cross effect matters.

4.3 Large number of entitites

Finally, let us discuss the case of a large number n of similar entities. If nis large, and the entities of similar sizes, we deduce from the Law of LargeNumbers (LLN) that the idiosyncratic terms can be diversified, whereas thesystematic factors cannot be. For expository purpose, let us consider a singlelinear factor model with γi = 1,∀i. We have :

X = (n∑

i=1

βi)f +n∑

i=1

ui.

Let us assume that the beta coefficients are i.i.d. with a positive meanE(β) > 0, and are independent of factor f and idiosyncratic errors ui, i =1, . . . , n. Let us also assume that these errors are independent with zeromean E(ui) = 0. The contributions are equal to :

22

RνP (X,Xi) =

∫E[Xi|X = qα(X)]νP (dα)

=

∫E[Xi|X/n = qα(X/n)]νP (dα),

since the quantile function is homogenous of degree 19.

By the LLN, we deduce that :

limn→∞

(X/n) = E(β)f.

Thus, the idiosyncratic part has been diversified, whereas the effect ofsystematic risk persists asymptotically.

When n = ∞, we get :

limn→∞E[Xi|X = qα(X)] = limn→∞E[Xi|X/n = qα(X/n)]

= E[Xi|E(β)f = qα[E(β)f ]]

= E[βif + ui|E(β)f = qα[E(β)f ]

= E[βif |E(β)f = qα[E(β)f ]

= limn→∞E[βif |X = qα(X)]

= limn→∞E[Xs,i|X = qα(X)].

In this limiting case, the contribution for entity i and the contributionfor its systematic component coincide, that is, RνP (X,Xi) = RνP,s

(X,Xi).Moreover, it is equivalent to condition either on X, or on the factor summaryE(β)f , or on factor f itself.

The derivation above helps to understand the contribution for systematicrisk used in Acharya et alii (2011), Brownlees, Engle (2010), which underlies

9The distribution of X depends on size n, but this dependence is not indicated forexpository purpose.

23

the daily updated systematic risk ranking published by NYU Stern’s Volatil-ity Lab. [www.systemicrisisranking.stern.nyu.edu], defined by 10 E[Xi|X =qα(X)]. This definition is valid under the assumptions above, i.e. when theunsystematic component can be diversified and when factor f can be identi-fied to the market portfolio.

For n large, but finite, it is possible by applying granularity theory[Gagliardini, Gourieroux (2011)], to evaluate at order 1/n the difference be-tween E[Xs,i|X = qα(X)] = E[βif |X = qα(X)], and E[Xi|X = qα(X)].

5 Required Capital

As mentioned in the introduction, the current regulation defines the requiredcapital in two steps. A dated measure of individual risk is first computed.This measure generally features an erratic behavior. Then a partial smooth-ing is applied to derive the required capital.

This practice does not distinguish the systematic and unsystematic com-ponents of the risk. We consider this question in Section 5.1, when the datedrisk measures are time independent Euler allocations of a DRM global riskmeasure. Since the ADM is fixed, it does not depend on the distributionof systemic risk Xt and in particular is time and path independent. In thisframework, we explain why the systematic and unsystematic componentshave to be smoothed differently to avoid spurious procyclical effects. Thenin Section 5.2 we discuss how a part of this effect might be avoided by con-sidering an appropriate path dependent global risk measure and ADM.

5.1 Change of link function

The link function [see equation (1.1) for a typical example] is a crucial elementof the regulatory policy. Whereas the trigger parameter is a control variatefor more or less prudential policy, the smoothing can be used to decrease, orincrease the effects of cycles.

The recent financial crisis revealed important drawbacks of a link functionlike (1.1) :i) The trigger parameter k has been fixed, independently of the market envi-ronment. We would have expected a reduced value during a liquidity crisis.

10Their analysis concerns stock market. Xi is the capitalization in stock i, whereas Xis the market portfolio value.

24

ii) As already mentioned in the introduction, the link function implies tworegimes that are a smoothed and an unsmoothed regime, the latter one ap-pearing with a large increase of the risk of entity i. However, the consequencesare not the same if this risk increase is due to an idiosyncratic shock, or toa shock on a systematic factor. In the first situation, there is an additionaldemand of liquid asset by entity i, which can be easily satisfied by the mar-ket. In the second situation, there is the demand for liquid asset by severalentities together, which may force financial institutions to deliver at fire-saleprices, creates the deleveraging spiral, that is, selling assets to reduce thedebt, and accentuates the cycles and the crisis.

Clearly the drawback of the link function (1.1) is the lack of distinctionbetween systematic and unsystematic risks, that is, between the micro andmacro prudential approaches.

Let us assume that the dated individual risk contributions are coherentand with a time independent ADM. Then functions R(.) and R(X, .) aretime independent and the values R(Xt, Xi,t) are Point-In-Time (PIT) con-tributions. Note that these contributions are time dependent because theyshould be computed with the joint distribution of (Xt, Xi,t) conditional onthe available information at date t, in particular the past values of (Xt, Xi,t).Instead of a formula of the type :

RCi,t = max[R(Xt, Xi,t), kt1

60

59∑

h=0

R(Xt−h, Xi,t−h)], (5.1)

an improved formula has to separate the two types of risks, in order to takeinto account the fact that the systematic components are highly dependent,and has to apply the two regime formula to the unsystematic componentonly 11. An alternative to formula (5.1) might be :

RCi,t = max[R(Xt, Xi,u,t), ku,t1

60

59∑

h=0

R(Xt−h, Xi,u,t−h)]

+ ks,t1

H

H−1∑

h=0

R(Xt−h, Xs,i,t−h) ≡ RCui,t +RCs

i,t. (5.2)

11To simplify the discussion, we assume a linear factor model, therefore a zero crossterm : Rs,u = 0.

25

The first component concerning the unsystematic risk is sufficient to pe-nalize the risky investments specific to a given entity, while avoiding a toovolatile evolution of the associated required capital. The second componentis linear to account for the strong serial dependence between the systematiccomponents. The smoothing window H is introduced for another purpose,not for avoiding a too volatile required capital, but for obtaining a counter-cyclical effect. In this respect, the smoothing window for systematic risk hastoo be larger than the usual 3-month (i.e. H = 60), and able to cover asignificant part of the cycle (one year for instance). Finally, the parameterks,t should be dependent of the position within the cycle, that is smaller inthe bottom of the cycle, larger in its top, to avoid the spurious creation of aliquidity gap.

The required capital RCi,t has to be provided to the regulators in liquid,high rated assets. Intuitively the components RCs

i,t and RCui,t have to be

included in different accounts of the Central Bank : the unsystematic com-ponent should be in an account specific to entity i, but all contributions forsystematic risk might be mutualized at least at the country level. Indeed,this latter account will serve to insure the global system and a mutualizationis usual for an unfrequent catastrophic event. If an entity is close to failuredue to a systematic effect, the total reserve for systematic component shouldbe used to avoid the failure of the entity and also some potential failures ofthe other ones by contagion.

Even if the application of different link functions to the systematic andunsystematic risk components seems relevant, its implementation will en-counter the same difficulties than the stress testing. Indeed, the systematicfactors have to be defined in a same way for all the entities by the regulatorsand there are many common risk factors which can be considered.

5.2 Change of global risk measure

The usual regulatory approach distinguishes the underlying baseline alloca-tions and the required capital. The aim is to pass from PIT measures of thetype R(Xt, Xi,t), which depend on the dated distribution of (Xt, Xi,t), to theThrough The Cycle (TTC) measures RCi,t, which depend on the current andlagged dated distributions of risks. However, this two step approach lackscoherency. For instance, the additivity property satisfied by the baselinecontribution is not satisfied by the required capital, since the link functionis nonlinear. Is it possible to develop a one-step TTC coherent approach ?

26

A possible answer is to define more precisely the global risk functionRt(Xt) of date t in a regulatory perspective. Let us consider the frameworkof Section 4 with underlying factors driving the systematic risk. These fac-tors have to be known and observed ex-post by the regulators. Thus, theregulators have an augmented information set including both the currentand lagged values of the global L&P,X, and factors F . Their risk measuresshould not only take into account the conditional distribution of current levelof risk X, but also the comparison of this level with the position in the cycle,function of F . In other words, the assumption, that the global reserve R(.)is a time independent function of the current conditional distribution of riskXt only, is likely not appropriate in a regulatory perspective. This globalmeasure should depend on the joint conditional distribution of Ft, Xt, and abetter notation would be R(F,X). We have seen that the allocation problemis an hedging problem w.r.t. the global L&P X. Similarly, the choice of anappropriate level of global reserve is also an hedging problem, but w.r.t. to areal economic benchmark and not the problem of controlling the stand alonerisk of X only. Typically, for a mortgage portfolio the optimal amount ofmortgage to be distributed (and its allocation between corporates, house-holds and sovereigns) should depend on the real estate cycle. The globalreserve might be a conditional quantile qα(Q)(X), with a critical value α(Q)function of the distribution Q of factor F , or a conditional expectation likeRc(F,X) = E(X) + E([X − c(F )]+), that is, the global risk measure couldchange along the real estate cycle.

A similar remark can be done for the contributions. The result of Corol-lary 2.5 is still valid, but with an increased information set. More precisely,the ADM νP should now be replaced by an ADM νP∗, where P ∗ denotes thejoint distribution of (X,F ).

6 Concluding remarks

The aim of this paper was to survey and complete the current literature oncapital allocation in a regulatory perspective and with special attention tosystematic risk. The main message of this paper is to avoid a crude use ofa coherent risk measure such as a VaR, or an ES for computing the reserveboth at the individual and global (system) levels. More precisely,

i) An allocation problem is different from a risk measurement problem.

27

The contributions are contingent to the level of global risk and have to satisfysome basic axioms.

ii) The axioms are not sufficient to define a unique contribution, and itcould be important to distinguish the risk measure underlying the measureof systemic global risk (characterized by a distortion measure if a DRM isretained for the global risk measure) from the allocation distortion measureexplaining how to allocate the global reserve.

iii) The allocation by ADM is applicable to any type of global risk mea-sure, in particular not necessarily homogenous like in Euler’s approach, andis also able to disentangle marginal systematic and unsystematic risk com-ponents and cross effects.

iv) The coherent contributions, satisfying axioms A1, A2, A3, are notunique and the degrees of freedom left to the decision makers are clearlyidentified : they are the mapping associating R(X) to a distribution P of Xand the mapping associating to P an ADM measure νP satisfying R(X) =∫

qα(X)νP (dα).

v) If the regulation has a purpose of economic policy, that is, if the mone-tary policy is not enough for the Central Bank, a time independent coherentrisk measure for the global risk is likely not appropriate. The subadditivityor homogeneity axioms may be unappropriate in a regulatory perspective.The global risk measure has to be chosen in relation with the economic en-vironment, especially with the position in the business or real estate cycle.The regulator faces an hedging problem, not the problem of managing thestand alone systemic risk.

The main questions that are still to be solved are the following ones :

i) What is (are) the objective(s) of the regulator ? In particular, is he/shepartly in charge of economic policy ? What is seen in our paper is that thedefinition of the required capital is an important instrument to control thequantity of credits and its distribution among firms, households, real estates,but also the leveraging among banks... Several Central Banks have for officialobjective the control of inflation by means of a prime rate. It seems importantto debate of the control of the credit distribution and leveraging by means

28

of the required capital [see e.g. Hellwig (2010) for a polemical discussion ofthe role of regulation].

ii) Once the objective is well-defined, how to choose the global level ofreserve Rt(Ft, Xt), which will be likely different from a global VaR, or fromthe sum of VaR’s of the different entities ?

iii) Once the objective and global level of reserve are well-defined, how tochoose the Allocation Distortion Measure, that is, the way of allocating theglobal reserve between the entities, between systematic, unsystematic andcross components?

To summarize, at a time where the databases, the statistical tools, themarginal and hedging risk measures, are almost in place, the different possibleobjectives of the regulation have now to be debated and their consequences onthe real economy and on the required capitals to be evaluated and compared.

29

R E F E R E N C E S

Acerbi, C., and T., Tasche (2002) : ”On the Coherence of ExpectedShortfall”, Journal of Banking and Finance, 26, 1487-1503.

Acharya, V. Pedersen, L., Philippon, T., and M., Richardson (2010) :”Measuring Systemic Risk”, Technical Report, Department of Finance, NYU.

Adrian, T., and M., Brunnermeier (2009) : ”CoVaR”, Federal ReserveBank of New-York, [revised version of the Technical Report 348, in (2008)].

Artzner, P., Delbaen, F., Eber, J., and D., Heath (1999) : ”CoherentMeasures of Risk”, Mathematical Finance, 9, 203-228.

Bertsimas, D., Lauprete, G., and A., Samarov (2004) : ”Shortfall asa Risk Measure : Properties, Optimization and Applications”, Journal ofEconomic Dynamics and Control, 28, 1353-1381.

Brownlees, C., and R., Engle (2010) : ”Volatility, Correlation, and Tailsfor Systemic Risk Measurement”, DP Stern Business School.

Denault, M. (2001) : ”Coherent Allocation of Risk Capital”, Journal ofRisk, 4, 1-34.

Dhaene, J., Goovaerts, M., and R., Kaas (2003) : ”Economic CapitalAllocation Derived From Risk Measures”, North American Actuarial Journal,7, 1-16.

Furman, E., and R., Zitikis (2008) : ”Weighted Risk Capital Allocations”,Insurance : Mathematics and Economics, 43, 263-269.

Gagliardini, P., and C., Gourieroux (2011) : ”Granularity Theory withApplication to Finance and Insurance”, mimeo, 300p.

Garman, M. (1997) : ”Taking VaR to Pieces”, RISK, 10, October, p70-71.

Gerber, H. (1979) : ”An Introduction to Mathematical Risk Theory”,Univ.of Pennsylvania, Philadelphia.

30

Gourieroux, C., Laurent, J.P., and O., Scaillet (2000) : ”Sensitivity Anal-ysis of Value-at-Risk”, Journal of Empirical Finance, 7, 225-245.

Hallerbach, W. (2002) : ”Decomposing Portfolio Value-at-Risk : A Gen-eral Analysis”, Journal of Risk, 5, 1-18.

Hellwig, M. (2010) :”Capital Regulation After the Crisis : Business asUsual N”, DP, Max Planck Institute.

Hesselager, O., and U., Andersson, (2002) : ”Risk Sharing and CapitalAllocation”, Tryg Insurance DP.

Kalkbrener, M. (2005) : ”An Axiomatic Approach to Capital Allocation”,Mathematical Finance, 15, 425-437.

Kim, J. (2010) : ”Measuring and Managing Systemic Risk”, Univ. ofWaterloo DP.

Koyluoglu, H., and J., Stoker (2002) : ”Honour Your Contributions”,Risk, 15, 90-94.

Kurth, A., and D., Tasche (2003) : ”Credit Risk Contributions to Value-at-Risk and Expected Shortfall”, Risk, 16, March., 84-88.

Litterman, R. (1996) : ”Hot Spots and Hedges”, The Journal of PortfolioManagement, 22, 52-75.

McNeil, A., Frey, R., and P., Embrechts (2005) : ”Quantitative RiskManagement”, Princeton University Press.

Patrick, G., Bernegger, S., and M., Ruegg (1999) : ”The Use of RiskAdjusted Capital to Support Business Decision Making”, Swiss ReinsuranceCompany DP.

Rosen, D., and D., Saunders (2010) : ”Risk Factor Contributions in Port-folio Credit Risk Models”, Journal of Banking and Finance, 34, 336-349.

Rothschild, M., and J., Stiglitz (1970) : ”Increasing Risk I : A Definition”,Journal of Economic Theory”, 2, 225-243.

31

Rudin, W. (1966) : ”Real and Complex Analysis”, Mc Graw Hill.

Shapley, L. (1953) : ”A Value for n-person Games”, in Contributionsto the Theory of Games, Volume II, by H., Kuhn and A., Tucker, editors,Annals of Mathematical Studies, 28, 307-317.

Sherris, M. (2007) : ”Risk Based Capital and Capital Allocation in In-surance”, Australian School of Business DP.

Tarashev, N., Borio, C., and K., Tsatsaronis (2009) : ”Allocating System-atic Risks to Individual Institutions : Methodology and Policy Application”,DP, BIS. Quartely Review.

Tasche, D. (2000) : ”Risk Contributions and Performance Measurement”,DP University of Munich.

Tasche, D. (2008) : ”Capital Allocation to Business Units and Sub-Portfolio : the Euler Principle”, in A. Resti, ed., Pillar II in the New BaselAccord : The Challenge of Economic Capital, p423-453, Incisive FinancialPublishing Lloyds TSB Bank, DP.

Uryasev, S., Theiler, U., and G., Serrano (2010) : ”Risk Return Opti-mization with Different Risk-Aggregation Strategies”, The Journal of RiskFinance, 11, 129-146.

Wang, S. (2000) : ”A Class of Distortion Operators for Pricing Financialand Insurance Risks”, Journal of Risk and Insurance, 67, 15-36.

Wang, S., and V., Young (1998) : ”Ordering Risks : Expected UtilityTheory Versus Yaari’s Dual Theory of Risk”, Insurance : Mathematics andEconomics, 22, 145-161.

32

A P P E N D I X 1

Proof of Proposition 2.2

Equivalence between i) and ii)

First note that the function :

y → U(y, x− y),

where U is concave is itself concave. Therefore ii) implies i).

Conversely, i) implies ii). Indeed let us denote by a a value for whichthe one-dimensional concave function U is maximal. If a is finite, U can bewritten as :

U(y) = U1(y) + U2(y) + U(a),

with U1(y) = 0, if y ≤ a,U(y)− U(a), otherwise,

U2(y) = U(y)− U(a), if y ≤ a,0, otherwise,

U1 (resp. U2) is a decreasing concave (resp. increasing concave) function.The result is deduced by noting that the function :

U(y1, y2) = U1(y1) + U2(x− y2) + U(a),

is concave, and that

U(y) = U(y, x− y)

If a = +∞ [resp. −∞], the result is deduced by noting that the function :(y1, y2) → U(y1) [resp. (y1, y2) → U(x− y2)] is concave.

Equivalence between ii) and iii)

This equivalence is given in Rothschild, Stiglitz (1970), Theorem 2.

QED

33

A P P E N D I X 2

Proof of Proposition 2.4

The proof is based on two Lemmas.

Lemma A.1 : The contributions : RµP(X,Xi) =

∫E[Xi|X = x]µP (dx)

satisfy the three axioms, if

∫xdµP (dx) = R(X).

Proof : The decentralization axiom is clearly satisfied and we consider belowthe two other axioms. i) We have :

n∑i=1

RµP(X,Xi) =

n∑i=1

∫E[Xi|X = x]µP (dx)

=

∫xdµP (dx),

= R(X),

which proves the additivity.

ii) Moreover if X∗1 ºX X1, we have :

E[X1|X = x]

= E[X∗1 + Z|X = x], with E(Z|X∗

1 , X) = 0, by Proposition 3.2 iii),

= E{E[X∗1 + Z|X∗

1 , X]|X = x}

= E[X∗1 |X = x], .

Thus, RµP(X,X∗

1 ) = RµP(X,X1).

We deduce the compatibility with the risk preordering, at least in a widesense.

QED

34

Lemma A.2 : Under the conditions of Proposition 2.4, all the contribu-tions are necessarily of the form given in Lemma A.1.

Proof : In the Hilbert space L2(Y ), the continuous linear form are neces-sarily of the type :

R(X,Xi) = E[a(Y )Xi],

where a(Y ) ∈ L2(Y ) [see e.g. Rudin (1966), Chapter 4]. The decentralizationand additivity axioms imply that a(Y ) can be written aP (X).

Moreover, we have :

E[aP (X)Xi] = E[aP (X)E(Xi|X)]

=

∫E[Xi|X = x]aP (x)P (dx),

where P (dx) is the distribution of global risk X. This expression is of theform given in Lemma A.1.

QED

Since aP is a measure density, R(X,Xi) is simply a weighted risk alloca-tion in the terminology of Furman, Zitikis (2008). When aP is positive, thecontribution can be interpreted as the value of Xi obtained by applying thepricing operator aP function of the distribution of X. This corresponds tothe premium calculation principle introduced in Gerber (1979).

35

A P P E N D I X 3

Explicit expression of the marginal expected shortfall

The proof is based on a succession of Lemmas

We have to prove that :

∂E[βX + Y |βX + Y > qα(β)]

∂β= E[X|βX + Y > qα(β)], (a.1)

where P [βX + Y > qα(β)] = 1− α, ∀β, (a.2)

and qα(β) = qα(βX + Y ), say.

Let us assume that the joint distribution of (X,Y ) is continuous withprobability density function 12 f(x, y). Equality (a.2) can be written as :

∫ [∫ ∞

−βx+qα(β)

f(x, y)dy

]dx = 1− α, ∀β.

Thus, by differentiating with respect to β, we get :

∫[x− ∂qα(β)

∂β]f [x, qα(β)− βx]dx = 0,∀β, (a.3)

which implies∂qα(β)

∂β= E[X|βX + Y = qα(β)].

The expected shortfall for βX + Y is :

ES(β)

= E[βX + Y |βX + Y > qα(β)]

=1

1− α

∫ [∫ ∞

−βx+qα(β)

(βx+ y)f(x, y)dy

]dx

Its derivative with respect to β is equal to :

12The general proof valid for any type of joint distribution for (X,Y ) has been given inTasche (2000), Lemma (5.6).

36

∂ES(β)

∂β=

1

1− α

∫ [∫ ∞

−βx+qα(β)

xf(x, y)dy

]dx

+1

1− α

∫ [x− ∂qα(β)

∂β

]qα(β)f [x, qα(β)− βx]dx

=1

1− α

∫ [∫ ∞

−βx+qα(β)

xf(x, y)dy

]dx [from (a.3)]

= E[X|βX + Y > qα(β)],

with is equation (a.1).

37